陈天桥携MiroThinker 1.5开年登场：跑赢万亿模型，实现小模型大智能

Core Insights - MiroMind team has launched MiroThinker 1.5, a flagship search intelligence model, which emphasizes "discovery-based intelligence" as a path to true general artificial intelligence [2][3] - The model aims to reconstruct understanding of the world under unknown conditions, focusing on research, verification, and correction rather than sheer data accumulation [2] Model Performance - MiroThinker 1.5 operates with 30 billion parameters, achieving performance comparable to larger models with 1 trillion parameters, demonstrating a high efficiency-to-intelligence ratio [3] - The model's cost per call is as low as $0.07, which is 1/20th of the cost of its competitor Kimi-K2-Thinking, while also providing faster inference [5] Interactive Scaling Concept - MiroThinker introduces "Interactive Scaling," shifting the focus from internal parameter expansion to external information interaction, enhancing reasoning capabilities [6][9] - The model is designed to function like a "scientist," emphasizing verification and correction over memorization, thus avoiding the pitfalls of traditional large models [8][10] Training Mechanism - The training process incorporates a "reason-verify-correct" loop, allowing the model to engage with external data for validation, which helps mitigate logical errors [9][12] - MiroThinker employs a time-sensitive training mechanism that restricts the model to only interact with information available before a given timestamp, ensuring realistic decision-making [12] Verification and Correction - The model encourages breaking down key judgments into verifiable sub-hypotheses and actively seeking external evidence, making the evidence-gathering process the primary training goal [11] - It emphasizes iterative verification, where reasoning is treated as a revisable process, allowing for adjustments based on conflicting evidence [11]